Harness Autonomy Tiers (HAT)
Grant autonomy at a pace you can defend.
Autonomous execution does not require granting complete authority on day one. The Plexus agent harness for Windows runs under a tier; your organization selects the Harness Autonomy Tier (HAT) appropriate for each workload, progressing from fully supervised toward autonomy within policy as the audit trail earns trust. Where the system runs is a separate choice, covered below.
HAT 01
Supervised
Operators approve actions before execution.
The agent prepares work and proposes the operational step. A human operator reviews and approves before any action executes. The substrate records that the check was raised: its category, its time, and the operator session attached to the agent. The operator's approval itself is kept by the harness, not the substrate.
Who it is for
Initial deployments and high-scrutiny workloads where operational trust is actively accumulating.
HAT 02
Coordinated
Agents coordinate and hand off within policy.
Multiple agents coordinate across systems and tasks, transferring context to other agents and human operators. Every handoff is attributed and cryptographically recorded, linked directly to the executing agent.
Who it is for
Multi-domain operations under active governance, including incident response, artifact assembly, and workflows spanning multiple enterprise systems.
HAT 03
Autonomous-within-Policy
Agents act autonomously within bounded policy.
The agent executes under its own authority within programmatic policy boundaries. Human operators engage on exceptions and policy violations rather than routine steps.
Who it is for
Scaled operations where the audit trail has established measurable reliability and per-step manual approval creates an operational bottleneck.
Deployment
Where it runs is a separate choice.
The tiers above describe who approves an action. Deployment describes what leaves your perimeter. The two are separate decisions, and each needs its own answer.
Today
Connected
The record substrate installs inside your perimeter and makes no outbound call to us. The harness sends a picture of the screen, and the text of the application in front, to the model you choose. By default that is a model provider's API, under your agreement with that provider.
On the roadmap
Air-gapped
The substrate is designed to run air-gapped; the harness is not air-gapped today, and we say so. Complete sovereignty, with the underlying models running entirely within your perimeter on dedicated hardware, is the milestone opened by custom model pipeline engagements.
The throughline
One audit substrate. Three levels of trust.
The HAT framework does not represent three separate products. It is a single unified substrate, configured to match the level of autonomy your risk policy is prepared to grant. The variable is the placement of human oversight.
The autonomy ladder provides the structured path to adoption, and the audit trail provides the evidence to climb.
Pre-launch · first design partners